SynRM-PINN

中文说明

Official PyTorch implementation and reproducibility package for:

High-Precision Standstill Flux Linkage Identification for Synchronous Reluctance Motors Based on Gradient Projection and Implicit Inductance Modeling

The repository provides the compact implicit-secant-inductance network, the multi-physics loss implementation, pairwise PCGrad, processed standstill hysteresis-injection datasets, simulation error datasets, external benchmark caches, representative checkpoints, and scripts used for the main experiments.

The manuscript is under review. A DOI, volume, issue, and page range have not been assigned.

Method summary

The proposed model combines:

• an implicit secant-inductance architecture that enforces zero-flux and current-parity properties by construction;
• data fitting with Maxwell reciprocity and discrete d/q voltage-equation residuals;
• pairwise Projecting Conflicting Gradients (PCGrad) across the nonzero weighted objectives;
• complete-cycle standstill data windows and early stopping based on the learning-rate threshold.

The common paper setting is:

Item	Value
Hidden layers	[6, 4]
Batch size	128
Maximum epochs	1000
Initial learning rate	1e-2
Early-stop learning rate	1e-5
MSE weight	1.0
Reciprocity weight	0.5
PDE weight	0.1
PCGrad mode	pairwise

Public-release scope

Included:

• processed experimental training CSV files used in the paper;
• filtered_gt_cache.npz, the constant-speed external benchmark used only for evaluation;
• processed simulation error datasets and simulation_gt_cache.npz;
• representative seed-21 checkpoints for the principal and plotted models;
• final numerical tables and selected SVG figures;
• scripts for the main multi-seed, ablation, dataset, error-propagation, and architecture studies.

Not included:

• original hardware acquisition records;
• the separate torque/rotor-angle diagnostic record;
• DSP control source code or compiled artifacts;
• Simulink models and generated files;
• manuscript drafts, reviewer correspondence, and internal revision reports;
• all repeated-run checkpoints and epoch-by-epoch logs.

See docs/DATA.md for the exact data boundary.

Installation

Python 3.11 is the reference environment. CPU and CUDA execution are both supported.

git clone https://github.com/KrisCCeng/SynRM-PINN.git
cd SynRM-PINN
python -m venv .venv

Windows PowerShell:

.venv\Scripts\Activate.ps1
python -m pip install --upgrade pip
pip install -r requirements.txt

Linux/macOS:

source .venv/bin/activate
python -m pip install --upgrade pip
pip install -r requirements.txt

PyTorch selects CUDA when an available CUDA build and GPU are detected; otherwise the code runs on CPU.

Validate the package

python scripts/validate_release.py
python -m unittest discover -s tests

The validator checks benchmark shapes, processed CSV schemas and finite values, representative checkpoint metadata, and the absence of restricted hardware or Simulink assets.

Quick start

Evaluate the released checkpoints

python evaluate.py --config configs/evaluate_pretrained.yaml

This evaluates the ANN, SCN, proposed PINN, analytical flux-current model, and direct third-order polynomial model on the same external 40 x 40 benchmark. Metrics and plots are written under:

outputs/pretrained/D130-Q100-D110Q200-2C/evaluation_results/

Train the proposed model

python train.py --config configs/pinn.yaml
python evaluate.py --config configs/pinn.yaml

Short smoke test

python train.py --config configs/smoke.yaml
python evaluate.py --config configs/smoke.yaml --no-generate_plots --no-include_baselines

The smoke configuration runs only three epochs and is not a paper result.

Reproduce the main studies

The commands below generate YAML files under outputs/generated_configs/, models/logs under outputs/runs/, and regenerated summaries under results/generated/. These generated directories are intentionally ignored by Git.

# Main experimental benchmark, initial scans, ablations, and seed-21 checks
python scripts/run_experimental_studies.py --phase all

# Three-seed independent physics-weight sensitivity
python scripts/run_independent_weight_multiseed.py

# Three-seed 3 x 3 local interaction study
python scripts/run_local_weight_robustness.py

# Three-seed complete-cycle, voltage-replacement, and multi-window comparisons
python scripts/run_dataset_multiseed.py

# Simulation error propagation and clean-data architecture checks
python scripts/run_simulation_error_propagation.py

# Five paired seeds on C2_medium and experimental data for five architectures
python scripts/run_practical_network_size_multiseed.py

# Representative benchmark error maps from released checkpoints
python scripts/generate_representative_error_heatmaps.py

The complete matrix contains many training runs. It is not necessary to rerun the matrix to inspect the reported values; final tables and figures are already provided in results/. See docs/REPRODUCIBILITY.md for the experiment-to-file mapping and expected compute behavior.

Repository structure

SynRM-PINN/
|-- configs/                 # Final, pretrained-evaluation, and smoke YAML files
|-- data/
|   |-- processed/           # Processed experimental datasets and benchmark
|   `-- simulation/          # Processed simulation error datasets and benchmark
|-- docs/                    # Data and reproducibility documentation
|-- outputs/
|   `-- pretrained/          # Representative checkpoints only
|-- results/                 # Final tables and selected vector figures
|-- scripts/                 # Validation and experiment runners
|-- src/                     # Models, losses, PCGrad, data loading, visualization
|-- tests/                   # Focused PCGrad behavior tests
|-- train.py
`-- evaluate.py

Main reported result

On the external constant-speed benchmark, the proposed [6,4] PINN obtains a three-seed average MAE of 0.01248 +/- 0.00254 Wb. This benchmark is excluded from gradient training and is used only for evaluation. Numerical tables are listed in results/README.md.

Citation

The paper has not yet received its final bibliographic metadata. Until then, use the following provisional entry and update the journal fields after publication:

@article{Ye2026SynRMPINN,
  title   = {High-Precision Standstill Flux Linkage Identification for
             Synchronous Reluctance Motors Based on Gradient Projection and
             Implicit Inductance Modeling},
  author  = {Ye, Cheng and Jia, Yankai and Song, Yusheng and Guo, Changxing and
             Shen, Chuanwen},
  journal = {Manuscript under review},
  year    = {2026},
  note    = {Code and processed data: https://github.com/KrisCCeng/SynRM-PINN}
}

Licenses

• Source code: MIT License.
• Released data and numerical result tables: CC BY 4.0.

Please cite both the associated paper and this repository when reusing the processed datasets.